Research map / Large language models / Retrieval and knowledge
Interact-RAG: Reason and Interact with the Corpus, Beyond Black-Box Retrieval
Yulong Hui, Chao Chen, Zhihang Fu, Yihao Liu, Jieping Ye, Huanchen Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| give the agent interaction primitives instead of black-box semantic search (ablation) | Interact-RAG · 2Wiki / MuSiQue / PopQA | EM: 63.4 -> 69.6 (2Wiki), 30.1 -> 34.8 (MuSiQue), 50.2 -> 56.0 (PopQA) | Improved |
| add SFT on synthesized trajectories before RL (ablation) | Interact-RAG · 2Wiki / MuSiQue / PopQA | EM: 59.0 -> 69.6 (2Wiki), 26.4 -> 34.8 (MuSiQue), 52.2 -> 56.0 (PopQA) | Improved |
| add RL (GRPO) after SFT (ablation) | Interact-RAG · 2Wiki / MuSiQue / PopQA | EM: 65.2 -> 69.6 (2Wiki), 28.1 -> 34.8 (MuSiQue), 45.6 -> 56.0 (PopQA) | Improved |
| add Multi-faceted Retrieval primitives (ablation) | Interact-RAG · 2Wiki / MuSiQue / PopQA | EM: 66.0 -> 69.6 (2Wiki), 34.6 -> 34.8 (MuSiQue) | Improved |
| add Anchored Matching primitive (ablation) | Interact-RAG · 2Wiki / PopQA | EM: 66.3 -> 69.6 (2Wiki), 53.4 -> 56.0 (PopQA) | Improved |
| add Context Shaping primitives (ablation) | Interact-RAG · MuSiQue | EM: 33.6 -> 34.8 | Improved |
| add the reasoning-enhanced workflow (planner, reasoner, executor) over a single prompt (ablation) | Interact-RAG-Workflow (training-free) · 2Wiki / MuSiQue / PopQA | EM: 52.0 -> 60.1 (2Wiki), 21.8 -> 24.1 (MuSiQue), 40.0 -> 43.6 (PopQA) | Improved |
| add interaction primitives to the training-free workflow (ablation) | Interact-RAG-Workflow · 2Wiki / MuSiQue / PopQA | EM: 56.3 -> 60.1 (2Wiki), 18.8 -> 24.1 (MuSiQue), 38.7 -> 43.6 (PopQA) | Improved |
| use SFT then RL rather than RL only (ablation) | Interact-RAG training · sampled subset of six test sets | EM | Improved |
| replace agent-weighted fusion with RRF or CombMNZ (ablation) | Multi-faceted retrieval · 2Wiki / MuSiQue / PopQA | EM: RRF 68.7, CombMNZ 69.8 vs 69.6 on 2Wiki | No clear effect |
| apply Best-of-3 sampling to Search-R1 instead of Interact-RAG | Search-R1 · 2WikiMultiHopQA | EM: 50.9 -> 56.2 | Improved |
| use Interact-RAG instead of Search-R1 for inference cost | Interact-RAG-7B · 2WikiMultiHopQA | latency (s): 17.2 -> 36.4 | Hurt |
Closest papers by meaning
- Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- Process vs. Outcome Reward: Which is Better for Agentic RAG Reinforcement Learning (NeurIPS 2025 · similarity 0.91)
- C-3PO: Compact Plug-and-Play Proxy Optimization to Achieve Human-like Retrieval-Augmented Generation (ICML 2025 · similarity 0.90)
- DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.89)
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.89)
- RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards (ICLR 2025 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
